不使用model.fit()时如何关闭Keras训练轮次日志

问题描述 #

我目前正在用GradientTape训练深度Q网络（DQN），代码如下：

with tf.GradientTape() as tape:
    q_values_current_state_dqn = self.dqn_architecture(states)
    one_hot_actions = tf.keras.utils.to_categorical(actions, self.num_legal_actions, dtype=np.float32) # e.g. [[0,0,1,0],[1,0,0,0],...]
    q_values_current_state_dqn = tf.reduce_sum(tf.multiply(q_values_current_state_dqn, one_hot_actions), axis=1)
    error = q_values_current_state_dqn - target_q_values
    loss = tf.keras.losses.Huber()(target_q_values, q_values_current_state_dqn)
    dqn_architecture_gradients = tape.gradient(loss, self.dqn_architecture.trainable_variables) # Computes the gradient using operations recorded in context of this tape.
    self.dqn_architecture.optimizer.apply_gradients(zip(dqn_architecture_gradients, self.dqn_architecture.trainable_variables))  

但训练时会不断输出如下进度日志，希望关闭：

1/1 [==============================] - 0s 34ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 11ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 11ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 13ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 11ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 11ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 10ms/step
1/1 [==============================] - 0s 12ms/step

我知道用model.fit()时可以设置verbose=0关闭日志，但不清楚用GradientTape时该怎么操作。

解决方案 #

这些日志是Keras默认的交互式训练日志输出，你可以通过以下两种方式关闭：

1. 全局关闭交互式日志 #

在训练代码的开头添加一行代码，一次性关闭所有Keras的交互式日志输出：

tf.keras.utils.disable_interactive_logging()

这会关闭包括进度条、每步训练日志在内的所有交互式输出，后续所有模型的前向/反向传播都不会再打印这类日志。

2. 临时关闭日志（上下文管理器） #

如果你只想在GradientTape训练的这段代码中关闭日志，其他地方保留输出，可以使用上下文管理器临时禁用：

import tensorflow as tf

with tf.keras.utils.disable_interactive_logging():
    with tf.GradientTape() as tape:
        q_values_current_state_dqn = self.dqn_architecture(states)
        one_hot_actions = tf.keras.utils.to_categorical(actions, self.num_legal_actions, dtype=np.float32)
        q_values_current_state_dqn = tf.reduce_sum(tf.multiply(q_values_current_state_dqn, one_hot_actions), axis=1)
        error = q_values_current_state_dqn - target_q_values
        loss = tf.keras.losses.Huber()(target_q_values, q_values_current_state_dqn)
    dqn_architecture_gradients = tape.gradient(loss, self.dqn_architecture.trainable_variables)
    self.dqn_architecture.optimizer.apply_gradients(zip(dqn_architecture_gradients, self.dqn_architecture.trainable_variables))

注意：把optimizer.apply_gradients放在外层不影响，因为日志主要来自模型的前向传播调用。

内容的提问来源于stack exchange，提问作者Bryan